This project implements deep learning models to predict the rotation angle of captcha pictures for breaking rotated captcha challenges. It provides two models: RotNet based on ResNet50 and RotNetR using RegNet_Y_3_2GF as backbone with 128 class outputs for rotation regression. Trained on COCO 2017 unlabeled dataset, RotNetR achieves approximately 7 degree average prediction error and lower cross-domain loss compared to the baseline RotNet.